net/ipv4/inet_connection_sock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Support for INET connection oriented protocols.
   8  *
   9  * Authors:     See the TCP sources
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/jhash.h>
  14
  15 #include <net/inet_connection_sock.h>
  16 #include <net/inet_hashtables.h>
  17 #include <net/inet_timewait_sock.h>
  18 #include <net/ip.h>
  19 #include <net/route.h>
  20 #include <net/tcp_states.h>
  21 #include <net/xfrm.h>
  22 #include <net/tcp.h>
  23 #include <net/sock_reuseport.h>
  24 #include <net/addrconf.h>
  25
  26 #if IS_ENABLED(CONFIG_IPV6)
  27 /* match_sk*_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses
  28  *                              if IPv6 only, and any IPv4 addresses
  29  *                              if not IPv6 only
  30  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  31  *                              IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  32  *                              and 0.0.0.0 equals to 0.0.0.0 only
  33  */
  34 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  35                                  const struct in6_addr *sk2_rcv_saddr6,
  36                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  37                                  bool sk1_ipv6only, bool sk2_ipv6only,
  38                                  bool match_sk1_wildcard,
  39                                  bool match_sk2_wildcard)
  40 {
  41         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  42         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  43
  44         /* if both are mapped, treat as IPv4 */
  45         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  46                 if (!sk2_ipv6only) {
  47                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  48                                 return true;
  49                         return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  50                                 (match_sk2_wildcard && !sk2_rcv_saddr);
  51                 }
  52                 return false;
  53         }
  54
  55         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  56                 return true;
  57
  58         if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard &&
  59             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  60                 return true;
  61
  62         if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard &&
  63             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  64                 return true;
  65
  66         if (sk2_rcv_saddr6 &&
  67             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  68                 return true;
  69
  70         return false;
  71 }
  72 #endif
  73
  74 /* match_sk*_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  75  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  76  *                              0.0.0.0 only equals to 0.0.0.0
  77  */
  78 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  79                                  bool sk2_ipv6only, bool match_sk1_wildcard,
  80                                  bool match_sk2_wildcard)
  81 {
  82         if (!sk2_ipv6only) {
  83                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  84                         return true;
  85                 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  86                         (match_sk2_wildcard && !sk2_rcv_saddr);
  87         }
  88         return false;
  89 }
  90
  91 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  92                           bool match_wildcard)
  93 {
  94 #if IS_ENABLED(CONFIG_IPV6)
  95         if (sk->sk_family == AF_INET6)
  96                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
  97                                             inet6_rcv_saddr(sk2),
  98                                             sk->sk_rcv_saddr,
  99                                             sk2->sk_rcv_saddr,
 100                                             ipv6_only_sock(sk),
 101                                             ipv6_only_sock(sk2),
 102                                             match_wildcard,
 103                                             match_wildcard);
 104 #endif
 105         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 106                                     ipv6_only_sock(sk2), match_wildcard,
 107                                     match_wildcard);
 108 }
 109 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 110
 111 bool inet_rcv_saddr_any(const struct sock *sk)
 112 {
 113 #if IS_ENABLED(CONFIG_IPV6)
 114         if (sk->sk_family == AF_INET6)
 115                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 116 #endif
 117         return !sk->sk_rcv_saddr;
 118 }
 119
 120 void inet_get_local_port_range(struct net *net, int *low, int *high)
 121 {
 122         unsigned int seq;
 123
 124         do {
 125                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 126
 127                 *low = net->ipv4.ip_local_ports.range[0];
 128                 *high = net->ipv4.ip_local_ports.range[1];
 129         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 130 }
 131 EXPORT_SYMBOL(inet_get_local_port_range);
 132
 133 static int inet_csk_bind_conflict(const struct sock *sk,
 134                                   const struct inet_bind_bucket *tb,
 135                                   bool relax, bool reuseport_ok)
 136 {
 137         struct sock *sk2;
 138         bool reuse = sk->sk_reuse;
 139         bool reuseport = !!sk->sk_reuseport && reuseport_ok;
 140         kuid_t uid = sock_i_uid((struct sock *)sk);
 141
 142         /*
 143          * Unlike other sk lookup places we do not check
 144          * for sk_net here, since _all_ the socks listed
 145          * in tb->owners list belong to the same net - the
 146          * one this bucket belongs to.
 147          */
 148
 149         sk_for_each_bound(sk2, &tb->owners) {
 150                 if (sk != sk2 &&
 151                     (!sk->sk_bound_dev_if ||
 152                      !sk2->sk_bound_dev_if ||
 153                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 154                         if ((!reuse || !sk2->sk_reuse ||
 155                             sk2->sk_state == TCP_LISTEN) &&
 156                             (!reuseport || !sk2->sk_reuseport ||
 157                              rcu_access_pointer(sk->sk_reuseport_cb) ||
 158                              (sk2->sk_state != TCP_TIME_WAIT &&
 159                              !uid_eq(uid, sock_i_uid(sk2))))) {
 160                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 161                                         break;
 162                         }
 163                         if (!relax && reuse && sk2->sk_reuse &&
 164                             sk2->sk_state != TCP_LISTEN) {
 165                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 166                                         break;
 167                         }
 168                 }
 169         }
 170         return sk2 != NULL;
 171 }
 172
 173 /*
 174  * Find an open port number for the socket.  Returns with the
 175  * inet_bind_hashbucket lock held.
 176  */
 177 static struct inet_bind_hashbucket *
 178 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 179 {
 180         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 181         int port = 0;
 182         struct inet_bind_hashbucket *head;
 183         struct net *net = sock_net(sk);
 184         int i, low, high, attempt_half;
 185         struct inet_bind_bucket *tb;
 186         u32 remaining, offset;
 187         int l3mdev;
 188
 189         l3mdev = inet_sk_bound_l3mdev(sk);
 190         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 191 other_half_scan:
 192         inet_get_local_port_range(net, &low, &high);
 193         high++; /* [32768, 60999] -> [32768, 61000[ */
 194         if (high - low < 4)
 195                 attempt_half = 0;
 196         if (attempt_half) {
 197                 int half = low + (((high - low) >> 2) << 1);
 198
 199                 if (attempt_half == 1)
 200                         high = half;
 201                 else
 202                         low = half;
 203         }
 204         remaining = high - low;
 205         if (likely(remaining > 1))
 206                 remaining &= ~1U;
 207
 208         offset = prandom_u32() % remaining;
 209         /* __inet_hash_connect() favors ports having @low parity
 210          * We do the opposite to not pollute connect() users.
 211          */
 212         offset |= 1U;
 213
 214 other_parity_scan:
 215         port = low + offset;
 216         for (i = 0; i < remaining; i += 2, port += 2) {
 217                 if (unlikely(port >= high))
 218                         port -= remaining;
 219                 if (inet_is_local_reserved_port(net, port))
 220                         continue;
 221                 head = &hinfo->bhash[inet_bhashfn(net, port,
 222                                                   hinfo->bhash_size)];
 223                 spin_lock_bh(&head->lock);
 224                 inet_bind_bucket_for_each(tb, &head->chain)
 225                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 226                             tb->port == port) {
 227                                 if (!inet_csk_bind_conflict(sk, tb, false, false))
 228                                         goto success;
 229                                 goto next_port;
 230                         }
 231                 tb = NULL;
 232                 goto success;
 233 next_port:
 234                 spin_unlock_bh(&head->lock);
 235                 cond_resched();
 236         }
 237
 238         offset--;
 239         if (!(offset & 1))
 240                 goto other_parity_scan;
 241
 242         if (attempt_half == 1) {
 243                 /* OK we now try the upper half of the range */
 244                 attempt_half = 2;
 245                 goto other_half_scan;
 246         }
 247         return NULL;
 248 success:
 249         *port_ret = port;
 250         *tb_ret = tb;
 251         return head;
 252 }
 253
 254 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 255                                      struct sock *sk)
 256 {
 257         kuid_t uid = sock_i_uid(sk);
 258
 259         if (tb->fastreuseport <= 0)
 260                 return 0;
 261         if (!sk->sk_reuseport)
 262                 return 0;
 263         if (rcu_access_pointer(sk->sk_reuseport_cb))
 264                 return 0;
 265         if (!uid_eq(tb->fastuid, uid))
 266                 return 0;
 267         /* We only need to check the rcv_saddr if this tb was once marked
 268          * without fastreuseport and then was reset, as we can only know that
 269          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 270          * owners list.
 271          */
 272         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 273                 return 1;
 274 #if IS_ENABLED(CONFIG_IPV6)
 275         if (tb->fast_sk_family == AF_INET6)
 276                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 277                                             inet6_rcv_saddr(sk),
 278                                             tb->fast_rcv_saddr,
 279                                             sk->sk_rcv_saddr,
 280                                             tb->fast_ipv6_only,
 281                                             ipv6_only_sock(sk), true, false);
 282 #endif
 283         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 284                                     ipv6_only_sock(sk), true, false);
 285 }
 286
 287 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 288                                struct sock *sk)
 289 {
 290         kuid_t uid = sock_i_uid(sk);
 291         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 292
 293         if (hlist_empty(&tb->owners)) {
 294                 tb->fastreuse = reuse;
 295                 if (sk->sk_reuseport) {
 296                         tb->fastreuseport = FASTREUSEPORT_ANY;
 297                         tb->fastuid = uid;
 298                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 299                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 300                         tb->fast_sk_family = sk->sk_family;
 301 #if IS_ENABLED(CONFIG_IPV6)
 302                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 303 #endif
 304                 } else {
 305                         tb->fastreuseport = 0;
 306                 }
 307         } else {
 308                 if (!reuse)
 309                         tb->fastreuse = 0;
 310                 if (sk->sk_reuseport) {
 311                         /* We didn't match or we don't have fastreuseport set on
 312                          * the tb, but we have sk_reuseport set on this socket
 313                          * and we know that there are no bind conflicts with
 314                          * this socket in this tb, so reset our tb's reuseport
 315                          * settings so that any subsequent sockets that match
 316                          * our current socket will be put on the fast path.
 317                          *
 318                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 319                          * do extra checking for all subsequent sk_reuseport
 320                          * socks.
 321                          */
 322                         if (!sk_reuseport_match(tb, sk)) {
 323                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 324                                 tb->fastuid = uid;
 325                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 326                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 327                                 tb->fast_sk_family = sk->sk_family;
 328 #if IS_ENABLED(CONFIG_IPV6)
 329                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 330 #endif
 331                         }
 332                 } else {
 333                         tb->fastreuseport = 0;
 334                 }
 335         }
 336 }
 337
 338 /* Obtain a reference to a local port for the given sock,
 339  * if snum is zero it means select any available local port.
 340  * We try to allocate an odd port (and leave even ports for connect())
 341  */
 342 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 343 {
 344         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 345         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 346         int ret = 1, port = snum;
 347         struct inet_bind_hashbucket *head;
 348         struct net *net = sock_net(sk);
 349         struct inet_bind_bucket *tb = NULL;
 350         int l3mdev;
 351
 352         l3mdev = inet_sk_bound_l3mdev(sk);
 353
 354         if (!port) {
 355                 head = inet_csk_find_open_port(sk, &tb, &port);
 356                 if (!head)
 357                         return ret;
 358                 if (!tb)
 359                         goto tb_not_found;
 360                 goto success;
 361         }
 362         head = &hinfo->bhash[inet_bhashfn(net, port,
 363                                           hinfo->bhash_size)];
 364         spin_lock_bh(&head->lock);
 365         inet_bind_bucket_for_each(tb, &head->chain)
 366                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 367                     tb->port == port)
 368                         goto tb_found;
 369 tb_not_found:
 370         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 371                                      net, head, port, l3mdev);
 372         if (!tb)
 373                 goto fail_unlock;
 374 tb_found:
 375         if (!hlist_empty(&tb->owners)) {
 376                 if (sk->sk_reuse == SK_FORCE_REUSE)
 377                         goto success;
 378
 379                 if ((tb->fastreuse > 0 && reuse) ||
 380                     sk_reuseport_match(tb, sk))
 381                         goto success;
 382                 if (inet_csk_bind_conflict(sk, tb, true, true))
 383                         goto fail_unlock;
 384         }
 385 success:
 386         inet_csk_update_fastreuse(tb, sk);
 387
 388         if (!inet_csk(sk)->icsk_bind_hash)
 389                 inet_bind_hash(sk, tb, port);
 390         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 391         ret = 0;
 392
 393 fail_unlock:
 394         spin_unlock_bh(&head->lock);
 395         return ret;
 396 }
 397 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 398
 399 /*
 400  * Wait for an incoming connection, avoid race conditions. This must be called
 401  * with the socket locked.
 402  */
 403 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 404 {
 405         struct inet_connection_sock *icsk = inet_csk(sk);
 406         DEFINE_WAIT(wait);
 407         int err;
 408
 409         /*
 410          * True wake-one mechanism for incoming connections: only
 411          * one process gets woken up, not the 'whole herd'.
 412          * Since we do not 'race & poll' for established sockets
 413          * anymore, the common case will execute the loop only once.
 414          *
 415          * Subtle issue: "add_wait_queue_exclusive()" will be added
 416          * after any current non-exclusive waiters, and we know that
 417          * it will always _stay_ after any new non-exclusive waiters
 418          * because all non-exclusive waiters are added at the
 419          * beginning of the wait-queue. As such, it's ok to "drop"
 420          * our exclusiveness temporarily when we get woken up without
 421          * having to remove and re-insert us on the wait queue.
 422          */
 423         for (;;) {
 424                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 425                                           TASK_INTERRUPTIBLE);
 426                 release_sock(sk);
 427                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 428                         timeo = schedule_timeout(timeo);
 429                 sched_annotate_sleep();
 430                 lock_sock(sk);
 431                 err = 0;
 432                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 433                         break;
 434                 err = -EINVAL;
 435                 if (sk->sk_state != TCP_LISTEN)
 436                         break;
 437                 err = sock_intr_errno(timeo);
 438                 if (signal_pending(current))
 439                         break;
 440                 err = -EAGAIN;
 441                 if (!timeo)
 442                         break;
 443         }
 444         finish_wait(sk_sleep(sk), &wait);
 445         return err;
 446 }
 447
 448 /*
 449  * This will accept the next outstanding connection.
 450  */
 451 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 452 {
 453         struct inet_connection_sock *icsk = inet_csk(sk);
 454         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 455         struct request_sock *req;
 456         struct sock *newsk;
 457         int error;
 458
 459         lock_sock(sk);
 460
 461         /* We need to make sure that this socket is listening,
 462          * and that it has something pending.
 463          */
 464         error = -EINVAL;
 465         if (sk->sk_state != TCP_LISTEN)
 466                 goto out_err;
 467
 468         /* Find already established connection */
 469         if (reqsk_queue_empty(queue)) {
 470                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 471
 472                 /* If this is a non blocking socket don't sleep */
 473                 error = -EAGAIN;
 474                 if (!timeo)
 475                         goto out_err;
 476
 477                 error = inet_csk_wait_for_connect(sk, timeo);
 478                 if (error)
 479                         goto out_err;
 480         }
 481         req = reqsk_queue_remove(queue, sk);
 482         newsk = req->sk;
 483
 484         if (sk->sk_protocol == IPPROTO_TCP &&
 485             tcp_rsk(req)->tfo_listener) {
 486                 spin_lock_bh(&queue->fastopenq.lock);
 487                 if (tcp_rsk(req)->tfo_listener) {
 488                         /* We are still waiting for the final ACK from 3WHS
 489                          * so can't free req now. Instead, we set req->sk to
 490                          * NULL to signify that the child socket is taken
 491                          * so reqsk_fastopen_remove() will free the req
 492                          * when 3WHS finishes (or is aborted).
 493                          */
 494                         req->sk = NULL;
 495                         req = NULL;
 496                 }
 497                 spin_unlock_bh(&queue->fastopenq.lock);
 498         }
 499
 500 out:
 501         release_sock(sk);
 502         if (newsk && mem_cgroup_sockets_enabled) {
 503                 int amt;
 504
 505                 /* atomically get the memory usage, set and charge the
 506                  * newsk->sk_memcg.
 507                  */
 508                 lock_sock(newsk);
 509
 510                 /* The socket has not been accepted yet, no need to look at
 511                  * newsk->sk_wmem_queued.
 512                  */
 513                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 514                                    atomic_read(&newsk->sk_rmem_alloc));
 515                 mem_cgroup_sk_alloc(newsk);
 516                 if (newsk->sk_memcg && amt)
 517                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 518
 519                 release_sock(newsk);
 520         }
 521         if (req)
 522                 reqsk_put(req);
 523         return newsk;
 524 out_err:
 525         newsk = NULL;
 526         req = NULL;
 527         *err = error;
 528         goto out;
 529 }
 530 EXPORT_SYMBOL(inet_csk_accept);
 531
 532 /*
 533  * Using different timers for retransmit, delayed acks and probes
 534  * We may wish use just one timer maintaining a list of expire jiffies
 535  * to optimize.
 536  */
 537 void inet_csk_init_xmit_timers(struct sock *sk,
 538                                void (*retransmit_handler)(struct timer_list *t),
 539                                void (*delack_handler)(struct timer_list *t),
 540                                void (*keepalive_handler)(struct timer_list *t))
 541 {
 542         struct inet_connection_sock *icsk = inet_csk(sk);
 543
 544         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 545         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 546         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 547         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 548 }
 549 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 550
 551 void inet_csk_clear_xmit_timers(struct sock *sk)
 552 {
 553         struct inet_connection_sock *icsk = inet_csk(sk);
 554
 555         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 556
 557         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 558         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 559         sk_stop_timer(sk, &sk->sk_timer);
 560 }
 561 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 562
 563 void inet_csk_delete_keepalive_timer(struct sock *sk)
 564 {
 565         sk_stop_timer(sk, &sk->sk_timer);
 566 }
 567 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 568
 569 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 570 {
 571         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 572 }
 573 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 574
 575 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 576                                      struct flowi4 *fl4,
 577                                      const struct request_sock *req)
 578 {
 579         const struct inet_request_sock *ireq = inet_rsk(req);
 580         struct net *net = read_pnet(&ireq->ireq_net);
 581         struct ip_options_rcu *opt;
 582         struct rtable *rt;
 583
 584         rcu_read_lock();
 585         opt = rcu_dereference(ireq->ireq_opt);
 586
 587         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 588                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 589                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 590                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 591                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 592                            htons(ireq->ir_num), sk->sk_uid);
 593         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 594         rt = ip_route_output_flow(net, fl4, sk);
 595         if (IS_ERR(rt))
 596                 goto no_route;
 597         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 598                 goto route_err;
 599         rcu_read_unlock();
 600         return &rt->dst;
 601
 602 route_err:
 603         ip_rt_put(rt);
 604 no_route:
 605         rcu_read_unlock();
 606         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 607         return NULL;
 608 }
 609 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 610
 611 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 612                                             struct sock *newsk,
 613                                             const struct request_sock *req)
 614 {
 615         const struct inet_request_sock *ireq = inet_rsk(req);
 616         struct net *net = read_pnet(&ireq->ireq_net);
 617         struct inet_sock *newinet = inet_sk(newsk);
 618         struct ip_options_rcu *opt;
 619         struct flowi4 *fl4;
 620         struct rtable *rt;
 621
 622         opt = rcu_dereference(ireq->ireq_opt);
 623         fl4 = &newinet->cork.fl.u.ip4;
 624
 625         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 626                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 627                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 628                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 629                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 630                            htons(ireq->ir_num), sk->sk_uid);
 631         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 632         rt = ip_route_output_flow(net, fl4, sk);
 633         if (IS_ERR(rt))
 634                 goto no_route;
 635         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 636                 goto route_err;
 637         return &rt->dst;
 638
 639 route_err:
 640         ip_rt_put(rt);
 641 no_route:
 642         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 643         return NULL;
 644 }
 645 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 646
 647 #if IS_ENABLED(CONFIG_IPV6)
 648 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 649 #else
 650 #define AF_INET_FAMILY(fam) true
 651 #endif
 652
 653 /* Decide when to expire the request and when to resend SYN-ACK */
 654 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 655                                   const int max_retries,
 656                                   const u8 rskq_defer_accept,
 657                                   int *expire, int *resend)
 658 {
 659         if (!rskq_defer_accept) {
 660                 *expire = req->num_timeout >= thresh;
 661                 *resend = 1;
 662                 return;
 663         }
 664         *expire = req->num_timeout >= thresh &&
 665                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 666         /*
 667          * Do not resend while waiting for data after ACK,
 668          * start to resend on end of deferring period to give
 669          * last chance for data or ACK to create established socket.
 670          */
 671         *resend = !inet_rsk(req)->acked ||
 672                   req->num_timeout >= rskq_defer_accept - 1;
 673 }
 674
 675 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 676 {
 677         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 678
 679         if (!err)
 680                 req->num_retrans++;
 681         return err;
 682 }
 683 EXPORT_SYMBOL(inet_rtx_syn_ack);
 684
 685 /* return true if req was found in the ehash table */
 686 static bool reqsk_queue_unlink(struct request_sock *req)
 687 {
 688         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 689         bool found = false;
 690
 691         if (sk_hashed(req_to_sk(req))) {
 692                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 693
 694                 spin_lock(lock);
 695                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 696                 spin_unlock(lock);
 697         }
 698         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 699                 reqsk_put(req);
 700         return found;
 701 }
 702
 703 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 704 {
 705         bool unlinked = reqsk_queue_unlink(req);
 706
 707         if (unlinked) {
 708                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 709                 reqsk_put(req);
 710         }
 711         return unlinked;
 712 }
 713 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 714
 715 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 716 {
 717         inet_csk_reqsk_queue_drop(sk, req);
 718         reqsk_put(req);
 719 }
 720 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 721
 722 static void reqsk_timer_handler(struct timer_list *t)
 723 {
 724         struct request_sock *req = from_timer(req, t, rsk_timer);
 725         struct sock *sk_listener = req->rsk_listener;
 726         struct net *net = sock_net(sk_listener);
 727         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 728         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 729         int qlen, expire = 0, resend = 0;
 730         int max_retries, thresh;
 731         u8 defer_accept;
 732
 733         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 734                 goto drop;
 735
 736         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 737         thresh = max_retries;
 738         /* Normally all the openreqs are young and become mature
 739          * (i.e. converted to established socket) for first timeout.
 740          * If synack was not acknowledged for 1 second, it means
 741          * one of the following things: synack was lost, ack was lost,
 742          * rtt is high or nobody planned to ack (i.e. synflood).
 743          * When server is a bit loaded, queue is populated with old
 744          * open requests, reducing effective size of queue.
 745          * When server is well loaded, queue size reduces to zero
 746          * after several minutes of work. It is not synflood,
 747          * it is normal operation. The solution is pruning
 748          * too old entries overriding normal timeout, when
 749          * situation becomes dangerous.
 750          *
 751          * Essentially, we reserve half of room for young
 752          * embrions; and abort old ones without pity, if old
 753          * ones are about to clog our table.
 754          */
 755         qlen = reqsk_queue_len(queue);
 756         if ((qlen << 1) > max(8U, sk_listener->sk_max_ack_backlog)) {
 757                 int young = reqsk_queue_len_young(queue) << 1;
 758
 759                 while (thresh > 2) {
 760                         if (qlen < young)
 761                                 break;
 762                         thresh--;
 763                         young <<= 1;
 764                 }
 765         }
 766         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 767         if (defer_accept)
 768                 max_retries = defer_accept;
 769         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 770                        &expire, &resend);
 771         req->rsk_ops->syn_ack_timeout(req);
 772         if (!expire &&
 773             (!resend ||
 774              !inet_rtx_syn_ack(sk_listener, req) ||
 775              inet_rsk(req)->acked)) {
 776                 unsigned long timeo;
 777
 778                 if (req->num_timeout++ == 0)
 779                         atomic_dec(&queue->young);
 780                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 781                 mod_timer(&req->rsk_timer, jiffies + timeo);
 782                 return;
 783         }
 784 drop:
 785         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 786 }
 787
 788 static void reqsk_queue_hash_req(struct request_sock *req,
 789                                  unsigned long timeout)
 790 {
 791         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 792         mod_timer(&req->rsk_timer, jiffies + timeout);
 793
 794         inet_ehash_insert(req_to_sk(req), NULL, NULL);
 795         /* before letting lookups find us, make sure all req fields
 796          * are committed to memory and refcnt initialized.
 797          */
 798         smp_wmb();
 799         refcount_set(&req->rsk_refcnt, 2 + 1);
 800 }
 801
 802 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 803                                    unsigned long timeout)
 804 {
 805         reqsk_queue_hash_req(req, timeout);
 806         inet_csk_reqsk_queue_added(sk);
 807 }
 808 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 809
 810 /**
 811  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 812  *      @sk: the socket to clone
 813  *      @req: request_sock
 814  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 815  *
 816  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 817  */
 818 struct sock *inet_csk_clone_lock(const struct sock *sk,
 819                                  const struct request_sock *req,
 820                                  const gfp_t priority)
 821 {
 822         struct sock *newsk = sk_clone_lock(sk, priority);
 823
 824         if (newsk) {
 825                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 826
 827                 newsk->sk_wait_pending = 0;
 828                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 829                 newicsk->icsk_bind_hash = NULL;
 830
 831                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 832                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 833                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 834
 835                 /* listeners have SOCK_RCU_FREE, not the children */
 836                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 837
 838                 inet_sk(newsk)->mc_list = NULL;
 839
 840                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 841                 atomic64_set(&newsk->sk_cookie,
 842                              atomic64_read(&inet_rsk(req)->ir_cookie));
 843
 844                 newicsk->icsk_retransmits = 0;
 845                 newicsk->icsk_backoff     = 0;
 846                 newicsk->icsk_probes_out  = 0;
 847                 newicsk->icsk_probes_tstamp = 0;
 848
 849                 /* Deinitialize accept_queue to trap illegal accesses. */
 850                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 851
 852                 security_inet_csk_clone(newsk, req);
 853         }
 854         return newsk;
 855 }
 856 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 857
 858 /*
 859  * At this point, there should be no process reference to this
 860  * socket, and thus no user references at all.  Therefore we
 861  * can assume the socket waitqueue is inactive and nobody will
 862  * try to jump onto it.
 863  */
 864 void inet_csk_destroy_sock(struct sock *sk)
 865 {
 866         WARN_ON(sk->sk_state != TCP_CLOSE);
 867         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 868
 869         /* It cannot be in hash table! */
 870         WARN_ON(!sk_unhashed(sk));
 871
 872         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 873         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 874
 875         sk->sk_prot->destroy(sk);
 876
 877         sk_stream_kill_queues(sk);
 878
 879         xfrm_sk_free_policy(sk);
 880
 881         sk_refcnt_debug_release(sk);
 882
 883         percpu_counter_dec(sk->sk_prot->orphan_count);
 884
 885         sock_put(sk);
 886 }
 887 EXPORT_SYMBOL(inet_csk_destroy_sock);
 888
 889 /* This function allows to force a closure of a socket after the call to
 890  * tcp/dccp_create_openreq_child().
 891  */
 892 void inet_csk_prepare_forced_close(struct sock *sk)
 893         __releases(&sk->sk_lock.slock)
 894 {
 895         /* sk_clone_lock locked the socket and set refcnt to 2 */
 896         bh_unlock_sock(sk);
 897         sock_put(sk);
 898
 899         /* The below has to be done to allow calling inet_csk_destroy_sock */
 900         sock_set_flag(sk, SOCK_DEAD);
 901         percpu_counter_inc(sk->sk_prot->orphan_count);
 902         inet_sk(sk)->inet_num = 0;
 903 }
 904 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 905
 906 static int inet_ulp_can_listen(const struct sock *sk)
 907 {
 908         const struct inet_connection_sock *icsk = inet_csk(sk);
 909
 910         if (icsk->icsk_ulp_ops)
 911                 return -EINVAL;
 912
 913         return 0;
 914 }
 915
 916 int inet_csk_listen_start(struct sock *sk, int backlog)
 917 {
 918         struct inet_connection_sock *icsk = inet_csk(sk);
 919         struct inet_sock *inet = inet_sk(sk);
 920         int err;
 921
 922         err = inet_ulp_can_listen(sk);
 923         if (unlikely(err))
 924                 return err;
 925
 926         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 927
 928         sk->sk_ack_backlog = 0;
 929         inet_csk_delack_init(sk);
 930
 931         /* There is race window here: we announce ourselves listening,
 932          * but this transition is still not validated by get_port().
 933          * It is OK, because this socket enters to hash table only
 934          * after validation is complete.
 935          */
 936         err = -EADDRINUSE;
 937         inet_sk_state_store(sk, TCP_LISTEN);
 938         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 939                 inet->inet_sport = htons(inet->inet_num);
 940
 941                 sk_dst_reset(sk);
 942                 err = sk->sk_prot->hash(sk);
 943
 944                 if (likely(!err))
 945                         return 0;
 946         }
 947
 948         inet_sk_set_state(sk, TCP_CLOSE);
 949         return err;
 950 }
 951 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 952
 953 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 954                               struct sock *child)
 955 {
 956         sk->sk_prot->disconnect(child, O_NONBLOCK);
 957
 958         sock_orphan(child);
 959
 960         percpu_counter_inc(sk->sk_prot->orphan_count);
 961
 962         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 963                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
 964                 BUG_ON(sk != req->rsk_listener);
 965
 966                 /* Paranoid, to prevent race condition if
 967                  * an inbound pkt destined for child is
 968                  * blocked by sock lock in tcp_v4_rcv().
 969                  * Also to satisfy an assertion in
 970                  * tcp_v4_destroy_sock().
 971                  */
 972                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
 973         }
 974         inet_csk_destroy_sock(child);
 975 }
 976
 977 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 978                                       struct request_sock *req,
 979                                       struct sock *child)
 980 {
 981         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 982
 983         spin_lock(&queue->rskq_lock);
 984         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 985                 inet_child_forget(sk, req, child);
 986                 child = NULL;
 987         } else {
 988                 req->sk = child;
 989                 req->dl_next = NULL;
 990                 if (queue->rskq_accept_head == NULL)
 991                         WRITE_ONCE(queue->rskq_accept_head, req);
 992                 else
 993                         queue->rskq_accept_tail->dl_next = req;
 994                 queue->rskq_accept_tail = req;
 995                 sk_acceptq_added(sk);
 996         }
 997         spin_unlock(&queue->rskq_lock);
 998         return child;
 999 }
1000 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1001
1002 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1003                                          struct request_sock *req, bool own_req)
1004 {
1005         if (own_req) {
1006                 inet_csk_reqsk_queue_drop(sk, req);
1007                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
1008                 if (inet_csk_reqsk_queue_add(sk, req, child))
1009                         return child;
1010         }
1011         /* Too bad, another child took ownership of the request, undo. */
1012         bh_unlock_sock(child);
1013         sock_put(child);
1014         return NULL;
1015 }
1016 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1017
1018 /*
1019  *      This routine closes sockets which have been at least partially
1020  *      opened, but not yet accepted.
1021  */
1022 void inet_csk_listen_stop(struct sock *sk)
1023 {
1024         struct inet_connection_sock *icsk = inet_csk(sk);
1025         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1026         struct request_sock *next, *req;
1027
1028         /* Following specs, it would be better either to send FIN
1029          * (and enter FIN-WAIT-1, it is normal close)
1030          * or to send active reset (abort).
1031          * Certainly, it is pretty dangerous while synflood, but it is
1032          * bad justification for our negligence 8)
1033          * To be honest, we are not able to make either
1034          * of the variants now.                 --ANK
1035          */
1036         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1037                 struct sock *child = req->sk;
1038
1039                 local_bh_disable();
1040                 bh_lock_sock(child);
1041                 WARN_ON(sock_owned_by_user(child));
1042                 sock_hold(child);
1043
1044                 inet_child_forget(sk, req, child);
1045                 reqsk_put(req);
1046                 bh_unlock_sock(child);
1047                 local_bh_enable();
1048                 sock_put(child);
1049
1050                 cond_resched();
1051         }
1052         if (queue->fastopenq.rskq_rst_head) {
1053                 /* Free all the reqs queued in rskq_rst_head. */
1054                 spin_lock_bh(&queue->fastopenq.lock);
1055                 req = queue->fastopenq.rskq_rst_head;
1056                 queue->fastopenq.rskq_rst_head = NULL;
1057                 spin_unlock_bh(&queue->fastopenq.lock);
1058                 while (req != NULL) {
1059                         next = req->dl_next;
1060                         reqsk_put(req);
1061                         req = next;
1062                 }
1063         }
1064         WARN_ON_ONCE(sk->sk_ack_backlog);
1065 }
1066 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1067
1068 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1069 {
1070         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1071         const struct inet_sock *inet = inet_sk(sk);
1072
1073         sin->sin_family         = AF_INET;
1074         sin->sin_addr.s_addr    = inet->inet_daddr;
1075         sin->sin_port           = inet->inet_dport;
1076 }
1077 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1078
1079 #ifdef CONFIG_COMPAT
1080 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
1081                                char __user *optval, int __user *optlen)
1082 {
1083         const struct inet_connection_sock *icsk = inet_csk(sk);
1084
1085         if (icsk->icsk_af_ops->compat_getsockopt)
1086                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
1087                                                             optval, optlen);
1088         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
1089                                              optval, optlen);
1090 }
1091 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
1092
1093 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
1094                                char __user *optval, unsigned int optlen)
1095 {
1096         const struct inet_connection_sock *icsk = inet_csk(sk);
1097
1098         if (icsk->icsk_af_ops->compat_setsockopt)
1099                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
1100                                                             optval, optlen);
1101         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1102                                              optval, optlen);
1103 }
1104 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
1105 #endif
1106
1107 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1108 {
1109         const struct inet_sock *inet = inet_sk(sk);
1110         const struct ip_options_rcu *inet_opt;
1111         __be32 daddr = inet->inet_daddr;
1112         struct flowi4 *fl4;
1113         struct rtable *rt;
1114
1115         rcu_read_lock();
1116         inet_opt = rcu_dereference(inet->inet_opt);
1117         if (inet_opt && inet_opt->opt.srr)
1118                 daddr = inet_opt->opt.faddr;
1119         fl4 = &fl->u.ip4;
1120         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1121                                    inet->inet_saddr, inet->inet_dport,
1122                                    inet->inet_sport, sk->sk_protocol,
1123                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1124         if (IS_ERR(rt))
1125                 rt = NULL;
1126         if (rt)
1127                 sk_setup_caps(sk, &rt->dst);
1128         rcu_read_unlock();
1129
1130         return &rt->dst;
1131 }
1132
1133 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1134 {
1135         struct dst_entry *dst = __sk_dst_check(sk, 0);
1136         struct inet_sock *inet = inet_sk(sk);
1137
1138         if (!dst) {
1139                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1140                 if (!dst)
1141                         goto out;
1142         }
1143         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1144
1145         dst = __sk_dst_check(sk, 0);
1146         if (!dst)
1147                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1148 out:
1149         return dst;
1150 }
1151 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);